Ensemble simulation and probabilistic climate projections

Ensemble simulations involve producing a group of parallel model simulations to characterise uncertainty (see discussion papers in Collins and Knight, 2007). Ensembles can be made with the same model but different initial conditions to characterise the uncertainty associated with climate variability, or by several models to represent the impact of uncertainty about choice of model formulation. More detailed ensemble simulations may also vary model parameters systematically to explore the modelling uncertainty. Run time is a major constraint for these methods because, in principle, an impractically large number of simulations may be needed to provide a stable estimate of the uncertainty. One way around this problem is to make a limited number of simulations with complex climate models and then use the results to 'train' a simpler statistical model emulator (Rougier et al., 2007).

This was the approach taken in producing the UKCP09 projections introduced earlier. The UKCP09 data derives from ensemble simulations, which provide not only a single trajectory for each climate model output but also a probability distribution. The probabilistic projections reflect some of the uncertainty associated with the climate model used to produce most of the data by varying its parameter systematically over several hundred simulations. The simulations are conducted for a period including both historical and future dates, and the results weighted to reflect how well each ensemble member agrees with real historical observations. In addition, a limited number of simulation results from other climate models are incorporated to recognise, to some degree at least, the uncertainty apparent in the differences between climate models.

Fig. 19.3 Percentage changes in flood peaks in 2071-2100 relative to a 1961-90 baseline simulated by Kay et al. (2006) using regional climate model data to drive a rainfall runoff model. Five digit numbers are UK National Water Archive gauging station catalogue numbers, percentages give change for a 10 year (left) and 50 year (right) return period. Reprinted from Journal of Hydrology, Volume 318, Kay, A. L., Jones, R. G. and Reynard, N. S., RCM rainfall for UK flood frequency estimation. II. Climate change results, pages 163-172, Copyright (2006), with permission from Elsevier.

Fig. 19.3 Percentage changes in flood peaks in 2071-2100 relative to a 1961-90 baseline simulated by Kay et al. (2006) using regional climate model data to drive a rainfall runoff model. Five digit numbers are UK National Water Archive gauging station catalogue numbers, percentages give change for a 10 year (left) and 50 year (right) return period. Reprinted from Journal of Hydrology, Volume 318, Kay, A. L., Jones, R. G. and Reynard, N. S., RCM rainfall for UK flood frequency estimation. II. Climate change results, pages 163-172, Copyright (2006), with permission from Elsevier.

Within UKCP09, a weather generator has been provided to simulate daily or hourly time series of weather variables corresponding to the future scenarios at 5-km grids across the country (Kilsby et al., 2007). This includes a stochastic rainfall model of the type outlined in Section 9.8 with parameters based on historical data that are then modified according to the UKCP09 predictions of percentage changes in monthly precipitation.

The probabilistic projections are made for specific GHG emissions scenarios, but the likelihood of these scenarios is not known and so the probability distributions of the model results are entirely conditional on the choice of scenario. It is therefore useful for hydrologists to understand that any analysis of impacts based on the probabilistic data will not give a predictive probability, but rather an indication of some (not all) of the uncertainty about the modelled impacts for a given scenario.

0 0

Post a comment